Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MemRerank: Preference Memory for Personalized Product Reranking

Zhiyuan Peng, Xuyang Wu, Huaixiao Tou, Yi Fang, Yu Gong

Abstract

LLM-based shopping agents increasingly rely on long purchase histories and multi-turn interactions for personalization, yet naively appending raw history to prompts is often ineffective due to noise, length, and relevance mismatch. We propose MemRerank, a preference memory framework that distills user purchase history into concise, query-independent signals for personalized product reranking. To study this problem, we build an end-to-end benchmark and evaluation framework centered on an LLM-based \textbf{1-in-5} selection task, which measures both memory quality and downstream reranking utility. We further train the memory extractor with reinforcement learning (RL), using downstream reranking performance as supervision. Experiments with two LLM-based rerankers show that MemRerank consistently outperforms no-memory, raw-history, and off-the-shelf memory baselines, yielding up to \textbf{+10.61} absolute points in 1-in-5 accuracy. These results suggest that explicit preference memory is a practical and effective building block for personalization in agentic e-commerce systems.

MemRerank: Preference Memory for Personalized Product Reranking

Abstract

LLM-based shopping agents increasingly rely on long purchase histories and multi-turn interactions for personalization, yet naively appending raw history to prompts is often ineffective due to noise, length, and relevance mismatch. We propose MemRerank, a preference memory framework that distills user purchase history into concise, query-independent signals for personalized product reranking. To study this problem, we build an end-to-end benchmark and evaluation framework centered on an LLM-based \textbf{1-in-5} selection task, which measures both memory quality and downstream reranking utility. We further train the memory extractor with reinforcement learning (RL), using downstream reranking performance as supervision. Experiments with two LLM-based rerankers show that MemRerank consistently outperforms no-memory, raw-history, and off-the-shelf memory baselines, yielding up to \textbf{+10.61} absolute points in 1-in-5 accuracy. These results suggest that explicit preference memory is a practical and effective building block for personalization in agentic e-commerce systems.

Paper Structure

This paper contains 25 sections, 1 equation, 12 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. LLM-based Rerankers
  4. Memory in Agentic Systems
  5. Methodology
  6. Dataset Construction
  7. MemRerank Framework
  8. RL Training with GRPO
  9. Experimental Setup
  10. Baselines
  11. No memory.
  12. Product-context baselines.
  13. LLM-extracted memory baselines.
  14. External memory baselines.
  15. Our method.
  16. ...and 10 more sections

Figures (12)

  • Figure 1: Impact of different memory extraction prompts on downstream personalized product reranking accuracy. "Within-category" denotes extracting only within-category shopping preference memory, while "Within- and cross-category" denotes extracting both within-category and cross-category shopping preference memory. For each setting, v1 uses a fixed set of suggested aspects, v2 allows the extractor to freely discover aspects from purchase history, and v3 combines light aspect guidance with flexible aspect discovery and explicit evidence grounding. Blue bars show the base extractor without training, and red bars show our trained extractor.
  • Figure 2: Impact of purchase history length on downstream personalized product reranking accuracy without explicit memory extraction. The vertical axis varies the number of within-category products, and the horizontal axis varies the number of cross-category products.
  • Figure 3: Mem0 prompt
  • Figure 4: MR.Rec Preference Patterns Extraction Prompt
  • Figure 5: MR.Rec User Profile Extraction Prompt
  • ...and 7 more figures